Demucs

App in the BluixApps catalog

What it is

Demucs is the state-of-the-art music source separation tool by Meta AI (FAIR) — splits songs into separate stems: vocals, drums, bass, and other instruments. Used by music producers, video creators, karaoke makers, podcast cleaners.

The "isolate the vocals" / "remove vocals for karaoke" tool, open-source and self-hostable.

What it's for

  • Stem separation — extract vocals, drums, bass, other from any song
  • Karaoke creation — remove vocals from licensed tracks
  • Vocal isolation — acapella extraction for remix/study
  • Podcast cleanup — separate background music from speech
  • Sample mining — extract individual instruments
  • Educational use — teach instrument identification

Who it's for

  • Music producers (acapella extraction, remix workflows)
  • DJs preparing mashups
  • Video editors cleaning audio
  • Karaoke service builders processing tracks
  • Music educators demonstrating instrument parts
  • Podcasters separating background music from speech

Why teams pick Demucs over alternatives

  • Hybrid license — code MIT, weights research-only (commercial weights via separate license)
  • State-of-art quality in open stem separation
  • Hybrid Transformer model — superior to LSTM-based predecessors
  • Multiple model variants — quality vs speed tradeoffs
  • Active research — Meta AI's continuous improvements
  • Fast — 10-20× real-time on RTX 4090

Integrations

  • Gradio web UI (BluixApps custom)
  • CLI mode for batch
  • Pair with: MusicGen (Demucs-extract → MusicGen-generate replacement stems)
  • Pair with: WhisperX (vocal extract → transcribe lyrics)
  • Output: 4 stem WAV files

Notable users & community

  • 9k+ GitHub stars
  • Meta AI (FAIR) research backing
  • Used in commercial music tools (Karaoke services)
  • Featured in audio engineering workflows
  • Active research community

Tips & operations

  • Model variants:
    • htdemucs: best general quality (default)
    • htdemucs_ft: fine-tuned, slightly better vocals
    • htdemucs_6s: 6-stem (also guitar + piano separately)
    • mdx_extra: alternative, different artifacts
  • Two-stems mode: just vocals vs instrumental (karaoke)
  • Speed:
    • CPU: 3-5 min per 3-min song
    • GPU: 10-20 sec per 3-min song
  • VRAM: 6 GB GPU optimal
  • License: code MIT, weights research-only.
    • Commercial use: either retrain on royalty-free data
    • Or use community alternatives (Spleeter for free commercial)
    • Or commercial APIs (LALAL.AI, Moises)
  • Production batch: CLI for entire albums

What we ship in BluixApps

  • Docker (pytorch CUDA 12.4 + demucs pip-installed)
  • Custom Gradio UI with model + two-stems selectors
  • Persistent volumes: models cache, input, output (4 stems WAV)
  • Port 7879 mapped
  • Install report at /root/bluixapps/demucs.txt
  • Model variant guidance
  • Use case examples (karaoke, sample mining, podcast cleanup)
  • License caveat documented (commercial restrictions on Meta-provided weights)
  • Pairing suggestions (MusicGen for replacement stems)
  • GPU pre-flight check via bluixapps_ensure_nvidia_runtime
  • Backup hook covers models + outputs
Read this app's deep dive on bluix.app ↗

Get this app — pick a BluixApps plan

Same catalog. Scaling tenant isolation, white-label and support tier.

TierTenantsCatalogSupportWhite-labelMonthly
Stacks119 curated stacksStandard$19/moDetailDeploy
Starter10Full catalogStandard+$15–25/mo$49/moDetailDeploy
Pro25Full catalogPriority bugfix+$15–25/mo$149/moDetailDeploy
Growth100Full catalogPriority bugfix+$15–25/mo$349/moDetailDeploy
Scale500Full catalog7-day window+$15–25/mo$799/moDetailDeploy
EnterpriseUnlimitedFull catalogPriority 7-dayBundled$1,499/moDetailDeploy

Powered by WHMCompleteSolution